The advancement of AI technology and related industries has propelled the intelligent era while simultaneously introducing safety risks. However, existing research predominantly perceives AI safety risks through the lens of general social risk governance, adopting a “process-outcome” second-order prevention approach. This perspective underestimates the examination of risk agents and neglects the future implications of risk consequences. To address these limitations, this study introduces the metaphor of extreme safety risks in AI as “Chernobyl disasters” from a bottom-line thinking perspective, proposing a “agent-process-outcome-process” fourth-order avoidance framework that builds upon conventional second-order prevention. After elucidating the dual manifestations and generative logic of extreme AI safety risks, this paper outlines three foundational considerations for AI safety bottom-line design based on the fourth-order avoidance framework:1.The transparency benchmark, transitioning from “agent” to “process,” addresses post-incident accountability challenges and cognitive dilemmas in AI disasters; 2. The preventive benchmark, shifting from “process” to “outcome,” aims to establish bottom-line prevention and maintain normalized controllability of AI risks; 3. The blocking benchmark tackles the intergenerational transmission of AI risks. Guided by these benchmarks, corresponding institutional constructs are proposed: 1.For transparency: institutional designs for algorithmic black-box transparency, human-machine interaction transparency, and data flow transparency; 2. For prevention: institutional frameworks addressing safety thresholds, worst-case scenarios, and redundancy/fault tolerance; 3.For blocking: institutional mechanisms to curb intergenerational risks such as energy consumption, ethical debt, and societal deconstruction.